1
Using AI and Natural Language Processing to Uncover
Population Social Determinants of Health Factors
Session #183, February 13, 2019
Varun Gupta, IT Director, Analytics and Data Management, Mt. Sinai Medical Center
Trish Birch, SVP and Global Consulting Leader, Cognizant Healthcare
2
Conflict of Interest
Varun Gupta
Has no real or apparent conflicts of interest to report.
Patricia Birch, MBA
Has no real or apparent conflicts of interest to report.
3
Agenda
• Welcome & Introduction
• Learning Objectives
• Using AI and Natural Language Processing to Uncover Population
Social Determinants of Health (SDoH) Factors
• Improving Health Outcomes with SDoH
• Q&A
4
Learning Objectives
Discuss how social determinants of health help providers improve
clinical models, understand disease progression and identify
phenotypes
Apply methods for cost effectively and accurately extracting social
determinants of health (SDoH) data from rich content in unstructured
clinical notes
Describe best practices for defining and creating an initial ontology set
Explain process of training natural language processing algorithms to
accurately identify SDoH in patient record notes
5
Using Natural Language
Processing to Uncover
Population SDoH Factors
6
About Mount Sinai
8 Hospitals
1 Medical School
9 Ambulatory Surgical Centers
190 Remote Clinical & Administrative Sites
38,000 Employees
7,800 Physicians
6,340 Faculty Members
3300+ Beds
138 Operating Rooms
3+ million Outpatient Visits
500+ thousand ER Visits
130+ thousand Inpatient Visits
425,000 IT Service Desk Calls
7
Moving the Needle:
Population Health and Analytics
Enabling IT Initiatives
Establish interoperability across clinical systems
Onboard applications and technologies to drive collaboration
Enable ‘Data as a Service’
Acute Care ($$$)
Primary Care ($$)
Community
Services ($)
Fee for Service
Value Based Services
8
Social Determinants Identification
• Notes (Progress, admission, procedure,
consultation notes); discharge summaries
Clinical Trials Recruitment
• Identification of patient cohorts
Improved Clinical Models
• Disease progression, phenotypes identification
Revenue Cycle Management
• Identify labs, procedures, medications, etc.
Computer-Assisted Coding/ Risk Adjustments
• Billing for correct services
Determining Risk Scores
• Smoker who quits ïƒ reduced risk of heart failure
• Behavior
(smoking, substance abuse etc. )
• Education
• Medications
• Discharge Education
• Economics
• ( financial status, education,
transportation )
• Safety
• Support
• Cultural
• Lab Results
• Demographics
• Prior Admissions
• Clinical
( current diagnostics, family history )
Unstructured
Structured
Natural Language Processing in Healthcare
9
Technology Framework
• Robust data ingestion
• FHIR-based API
platform
• Hadoop data lake
• Hybrid cloud-based
architecture on Azure
• Unified data desk and
extended data teams
10
• Housing, Income
• Food and Nutrition
Economic
Stability
• Lack of education or Current
education level causing any
challenges
Education
• Medication, Prevention care
• Insurance, Language
barriers
Healthcare
System
• Transportation
• Concerns on overall safety
and well being
Physical
Environment
• Depression/anxiety
• Substance abuse and Sleep
relates problems
Behavioral
Health
• Patient’s care giving
responsibilities are having adverse
effect on Patient’s health
Care Giving
Responsibility
• Legal issues like divorce, child
custody, alimony etc.
Legal
• Formal support: health home etc.
• Informal support: Friends,
neighbors, relatives etc
Support
System
• How active patient is and does
patient have avenues to stay
active
Physical
Activity
• Includes, Identification of
measures (not) taken for fall
safety
Fall Safety
• Special health needs like vision,
hearing, and Urinary incontinence
Special Health
needs
Legend : Phase 1 Phase 2 Phase3
NLP Process – Categorization
11
NLP Workflow
Clinical Notes Extraction
Corpus Creation
Garbage Filtering
Lemmatization
Document Classification
Duplicate Removal
Collation of Notes
Spell Check
Synonym Replacement
Document Term Matrix
WORDS
Bag of Words
12
Model Outcomes
All Medicaid notes added to the database during Dec ’17 to Jul’18 were used for the
model execution. Separate sample outputs for each model were manually validated by
clinical SMEs to drive model parameters.
Economic
Stability
Education
Healthcare
System
Physical
Environment
97%
Accuracy
Sensitivity: 96%
Specificity: 99%
93%
Accuracy
Sensitivity: 97%
Specificity: 88%
89%
Accuracy
Sensitivity: 91%
Specificity: 85%
86%
Accuracy
Sensitivity: 98%
Specificity: 75%
13
Time Period
Medicaid
Aug
2016
2018
7,240,320
Encounters
226,368
Patients
SDoH Category
Patient
Count
% of
Patients with SDoH
Economic Stability 35,187
Healthcare 35,781
Education 8,770
Physical Environment 18,685
Total Patients with SDoH 70,541
16%
4%
9%
&
31% of total patients who had an encounter Aug 2016-July 2018 were
found to have at least one SDoH factor.
16%
Capturing Insights
to
Oct
14
Extraction of
Clinical notes &
Social determinants
survey data
EMR / Data
Repositories
Data Scientists
NLP Model & SDH
Repository at patient
level
Care Gives / Population
Health Managers
Industry standard
SDOH libraries
Unified Medical Language system
1. Economic Stability
2. Education
3. Physical environment
4. Health system
5. Behavioral Health
6. Caregiving responsibilities
7. Legal, Support system
8. Physical activity
Next Steps to Scale
15
• Creating an ontology unique to Mt. Sinai, but universal enough to effectively capture
SDoH
• Testing the accuracy of the algorithm’s output was labor-intensive.
• Integrating findings into clinical workflows and systems
• Integrating the generated outcome back to the individual patient EMR
Challenges
• Analyze your available data
• Identify opportunities: process fit, most impactful SDoH
• Start simply, but be comprehensive
• Develop a roadmap
Best Practices
16
Improving Health Outcomes
with SDoH
17
Healthcare to Social care
Medicare
Care, 10%
Biology
(Genetics),
30%
Physical
Environment,
5%
Social
Circumstances,
15%
Individual
Behavior,
40%
Why should you care?
What Impacts
Health
Outcomes?
1
Maricopa County
2 n = 305,381
• Higher utilization
• Increased presence of
chronic disease
• Higher spend
18
Opportunity Analysis:
Impact on Payers & Providers
Health Plans
• High Cost of Care
• Unpredictable Actuarial
Pool
• Member Satisfaction
Health Systems
• Increased cost of care and
financial risk
• Loss of Clinical time and
resources
• Alignment of new mission:
treating whole person and
community
19
Today’s SDoH Tools are Inefficient
Lack of Integration Across the Care Journey
• Manual Documentation
• Lack of Integration across the Care Journey
• Manual Entry into EMR
20
Lack of Integration Across Care Journey
Upon identification of a SDoH, there’s no efficient
way to communicate and integrate this information
across the care continuum
Inefficient
Integration
Poor management of referral resources leads to
difficulty identifying which resources will be most
effective and supportive for a patient’s needs
Referral
Resources
Care Managers and Patient Navigators are often
responsible for a handful of patients at once
Overburdened
System
21
Digital Tools can Uncover and Integrate
SDoH into Today’s Care Continuum
Creates efficiency during care touchpoints
Leverages pre-existing data
Allows for automatic integration into the
care continuum
22
SDoH: Whole Person Care
Medical
Excellence
Mental Health
& Wellness
Social Care
23
Caring for the Whole Person Requires
Transformation at all 3 Levels
Patient Health System Community
What can be done to better identify and integrate SDoH
into a patient’s care journey across all 3 levels?
24
SDoH Transformation Takeaways
1.
Solutions
must be
efficient
2.
Integration
across the
care journey is
needed
3.
A comprehensive
approach is
required
26
Appendix
27
Sample Data
